Yunnan Province is one of the provinces in China most frequently affected by debris flows, and mountain disaster risk is a major threat to regional ecological security and the sustainable development of human settlements. To address the tendency of existing susceptibility assessments to focus on single models and overlook elevational differentiation, this study selected 14 conditioning factors from multisource data, including topographic and meteorological data, and developed Optuna-optimized Random Forest (RF), extreme gradient boosting (XGBoost), and Stacking ensemble models. On the same 1702 randomly held-out test samples, the Stacking model achieved the highest AUC among the seven models evaluated. SHAP analysis showed that distance from ravines was the dominant factor across the study area, while DEM, NDVI, and mean annual precipitation were important supplementary factors. The susceptibility mapping results showed that high- and very-high-susceptibility zones were concentrated in tectonically active, deeply dissected areas of northwestern, northeastern, and southwestern Yunnan. The vertical zonation analysis showed marked differences in the distribution of high- and very-high-susceptibility zones among the three elevation zones, reflecting spatial differences in elevation, vegetation cover, and long-term precipitation conditions across the study area. The Stacking framework developed in this study combines predictive performance with interpretability, and the identified vertical differentiation patterns enrich the understanding of regional debris-flow distribution. The assessment results can provide scientific support for sustainable spatial planning in Yunnan Province, the targeted allocation of disaster-prevention resources, and the safety of mountain settlements, and have practical value for reducing regional disaster risk and promoting the long-term, coordinated sustainable development of ecological and socioeconomic systems.
Landslide assessment requires static factors defining baseline conditions and dynamic factors acting as triggers. However, integrating various factors into susceptibility assessments increases uncertainty in multi-criteria decision making (MCDM) frameworks and reduces the accuracy of data-driven models. This study proposes a novel landslide susceptibility assessment for Yunnan, China, integrating factor weights from MCDM and one-class classification. The approach advances susceptibility mapping by incorporating 18 static and dynamic factors, including variability in precipitation, soil moisture, vegetation, and land cover. Monte Carlo simulations validated the factor-weighting scheme of Best–Worst Method and Isolation Forest, identifying slope angle and precipitation (mean and variability) as dominant drivers, reflecting the combined influence of steep terrain and monsoonal rainfall in Yunnan. The susceptibility map reveals that 15 % of Yunnan falls into the very low, 27 % into the low, 28 % into the moderate, 22 % into the high, and 8 % into the very high susceptibility classes. This indicates that a substantial portion of the province remains highly susceptible. Susceptibility classes were examined using analysis of variance. Slope angle, mean precipitation, and soil moisture emerged as the most influential factors in distinguishing landslide-prone zones. These findings supports spatiotemporal adaptive planning and targeted mitigation strategies to reduce landslide risk in Yunnan.